Question Xeon-style Workstation CPUs - obsolete for non-enterprise?

[dsc106]

I understand why multi million dollar enterprises may opt for Xeons for certain certifications and familiarity. I am wondering if from a performance and reliability perspective if these types of CPUs are even relevant anymore for most of us? And, even for enterprise, how much merit there *actually* is to them over HEDT processors?

My understanding is:

- As multicore has expanded on HEDT, benefits of workstation decline (when most software can't even scale to as many threads as a single chip can do, multi-CPU support is extremely niche need)
- ECC memory reliability is such a minor improvement it only matters in server scenarios or really complex calculations
- More lanes and I/O hardly matters anymore with so much I/O available on Threadripper PCI-E 4 systems, and by virtue of how flexible and how much data even a single I/O port carries today
- More RAM support is becoming increasingly niche as we reach 128gb-256gb of RAM on desktop
- More memory channels (6 vs 4 on HEDT or 2 on Desktop) - but in this that big of a benefit or are the benefits miniscule?
- "Higher quality parts" seems like a misnomer - yes, high quality parts, but aren't the same quality parts available for desktop/HEDT if you buy the right components?
- Longevity - supposedly they can run 24/7 at full load for a longer lifetime, but I've never heard of anyones non-OC'ed CPU dying out? I mean, if I took a desktop or HEDT chip from Intel or AMD that was well cooled and just started it today with an endless supply of full load video encoding, when would it "die" statistically?

Bottom line, it seems that "workstation" used to mean "professional", but today a workstation CPU is a poor choice for almost any professional content creator. I don't mean a poor value choice, I mean if prices were equal, still the lesser choice. Am I missing something?

Does workstation CPU have a place outside server, enterprise, and long-form scientific calculations like quantum models and that sort of thing? And in large organizations/enterprise, are benefits of the platform becoming increasingly dubious and relying more on managers not wanting to stick there necks out to break status quo under the false impression that a workstation cpu will be "more reliable"?

Thanks! Just wondering what the 2020 assessment of this classic divide are.

 

[NTMBK]

ECC is a bigger deal than you make out. I don't like speccing out a work machine without it, it's saved my bacon in the past.

I had a system with motherboard problems; something to do with the traces between the two sockets in the end. Subtle problem that only manifested when I really hammered the machine with a big CPU load. Performance would suddenly slow to a crawl- turned out that the ECC memory was detecting and correcting thousands of errors. We got the logs, showed them to Dell, and got the board replaced the next day. Without ECC I would have had data silently corrupted.

 

[burninatortech4]

I agree that I don't see the Xeon workstation having much of a niche anymore. That's entirely Intel's fault. They have butchered their own product segmentation through a combination of the 10nm debacle and dumb marketing choices.

ECC is one of those things you wish you had AFTER you discovered data corruption. Even if it's not needed 99% of the time. It's all about that 1% for work you're getting paid to do or are responsible for. Might as well not roll the dice on a work machine.

If I had a Threadripper 3rd Gen machine or similar for work I would 100% use ECC.

 

[dsc106]

I’m confused, ECC can cause long term data corruption? I thought it would just cause crashes with active memory corruptions, but you’re saying it can corrupt files saved to the HDD/SSD?

 

[Markfw]

I’m confused, ECC can cause long term data corruption? I thought it would just cause crashes with active memory corruptions, but you’re saying it can corrupt files saved to the HDD/SSD?

I think you misunderstood the posts. I don't see that being said here. I see people saying that if they DON'T have ECC, and then find data corruption, at that time they wish they DID have ECC.

 

[dsc106]

No I understood, I mean, find data corruption where? Like things are corrupt actively while working? IE a long calculation?

Or as in, my video project file or word document gets corrupted because memory rewrote it to the HDD wrong?

 

[moinmoin]

No I understood, I mean, find data corruption where? Like things are corrupt actively while working? IE a long calculation?

Or as in, my video project file or word document gets corrupted because memory rewrote it to the HDD wrong?

Data corruption through hardware faults.

 

[dsc106]

I don’t know what you mean by that specifically. What kind of data corruption.

 

[dsc106]

Also, not to get us too far down this trail - is there anything beyond ECC? Or was everything I else I mentioned pretty much a ... “yup”... largely pointless? (Especially considering you can get ECC memory on non workstation boards like threadripper now)

 

[VirtualLarry]

Also, not to get us too far down this trail - is there anything beyond ECC? Or was everything I else I mentioned pretty much a ... “yup”... largely pointless? (Especially considering you can get ECC memory on non workstation boards like threadripper now)

When you say "beyond ECC", I'm going to presume that you mean, "beyond ECC RAM", because plenty of things use ECC technology. HDDs and SSDs do, for example.

There are also higher-level forms of ECC that you can use. PAR2 parity archives, often used for transmitting valuable content over unreliable networks, is a form of forward ECC. (As long as you have as many parity chunks as the missing/bad blocks of the original content, you can process it using the ECC algorithms, and recover the original data contents.)

 

[Atari2600]

I don’t know what you mean by that specifically. What kind of data corruption.

Within DRAM 1s become 0s and 0s become 1s. Which obviously is not the data you were expecting.

How it arrives is a different question.

No I understood, I mean, find data corruption where? Like things are corrupt actively while working? IE a long calculation?

Yes, this is the type of corruption being discussed. Live data being corrupted.

 

[lobz]

ECC is a bigger deal than you make out. I don't like speccing out a work machine without it, it's saved my bacon in the past.

I had a system with motherboard problems; something to do with the traces between the two sockets in the end. Subtle problem that only manifested when I really hammered the machine with a big CPU load. Performance would suddenly slow to a crawl- turned out that the ECC memory was detecting and correcting thousands of errors. We got the logs, showed them to Dell, and got the board replaced the next day. Without ECC I would have had data silently corrupted.

Don't count on the workstation threadrippers (1H 2020) lacking official ECC support.

 

[LikeLinus]

Also, not to get us too far down this trail - is there anything beyond ECC?

Support for more RAM
Support for multi-cpu
Support for larger cache.
Support for heavy-duty work cycles and 24/7 usage.

Both the RAM and Multi-CPU allow for far more Virtual Machines in a smaller space. Fewer machines to maintain. I'm sure there is more, but that is off the top of my head.

We use them for 3D modeling, editing machines (dual Xeon MacPros), motion graphics and render farms.

 

[Markfw]

Don't count on the workstation threadrippers (1H 2020) lacking official ECC support.

I believe they are already supported. We know the threadripper is, and here is what is on all of the ASRock trx40 motherboards:

Supports DDR4 4666(OC)+/ 4600(OC)/ 4533(OC)/ 4466(OC)/ 4400(OC)/ 4333(OC)/ 4266(OC)/ 4200(OC)/ 4133(OC)/ 4000(OC)/ 3866(OC)/ 3800(OC)/ 3733(OC)/ 3600(OC)/ 3466(OC)/ 3200/ 2933/ 2667/ 2400/ 2133 ECC & non-ECC, un-buffered memory (U-DIMM)

 

[dsc106]

Support for more RAM
Support for multi-cpu
Support for larger cache.
Support for heavy-duty work cycles and 24/7 usage.

Both the RAM and Multi-CPU allow for far more Virtual Machines in a smaller space. Fewer machines to maintain. I'm sure there is more, but that is off the top of my head.

We use them for 3D modeling, editing machines (dual Xeon MacPros), motion graphics and render farms.

Ok so on this, some of my questions on the possible obsolence for most:

- More RAM - programs that make meaningful use of more than 256gb of RAM (especially in this era of extreme speed SSDs) seems to make this sort of a non-issue but for all of the most extreme/rare/niche use cases?

- Multi CPU - 32 and now 64 cores in a single chip, which hardly anything can take advantage of, seem to make the hassles of multi CPU even more niche (and frankly, unneccessary for most) than they already work?

- Larger cache - vs the extremely large cache on Threadripper, does this have much of a meaningful performance impact for content creation?

- Heavy duty work cycles and 24/7 usage - I am looking for a lot more specificity here. What does this statistically mean? I have never heard of anyone breaking their properly cooled CPU. What's the 10 year breakdown of this. If I put a Threadripper or Ryzen, or Intel i9 on a 24/7 render cycle for the next 10 years would it die VS the Xeon? What are the stats/MTF/etc?

Taking this further, when you say you use this for 3D modeling, editing machines, motion graphics. Do those programs actually take meaningful advantage of dual CPUs? And, do the clockspeed/IPC advantages of Threadrippers negate that? I mean, in my looking around most software will work better with 16-32 cores at higher clock speeds than a Xeon at lower clock speeds, and not even scale well to dual CPU?

 

[dsc106]

Yes, this is the type of corruption being discussed. Live data being corrupted.

Ok so if it's just live data being corrupted in rare instances, for most content creators this is sort of a moot issue yeah? And those who want ECC can still get it on a non-Xeon board. So that's not a Xeon/workstation advantage anymore if my understanding is correct.

 

[Markfw]

Ok so if it's just live data being corrupted in rare instances, for most content creators this is sort of a moot issue yeah? And those who want ECC can still get it on a non-Xeon board. So that's not a Xeon/workstation advantage anymore if my understanding is correct.

ECC works on most any AMD SP3 and TR4 and TR4X motherboard and most any motherboard that supports Xeon or Intel server CPU's. Its definitely NOT a Xeon advantage.

But it is recommended for any system that is critical for work or a business and definitely required for all servers (realistically, not mandated by all hardware, except many Xeon and ALL EPYC systems)

 

[lobz]

I believe they are already supported. We know the threadripper is, and here is what is on all of the ASRock trx40 motherboards:

Supports DDR4 4666(OC)+/ 4600(OC)/ 4533(OC)/ 4466(OC)/ 4400(OC)/ 4333(OC)/ 4266(OC)/ 4200(OC)/ 4133(OC)/ 4000(OC)/ 3866(OC)/ 3800(OC)/ 3733(OC)/ 3600(OC)/ 3466(OC)/ 3200/ 2933/ 2667/ 2400/ 2133 ECC & non-ECC, un-buffered memory (U-DIMM)

Yes, but I meant officially and certificated, I believe it'll come with the W (or whatever the naming scheme will turn out to be) series, with a bit more conservative boost clocks and 8 channel memory and ECC support.

 

[dsc106]

ECC... is recommended for any system that is critical for work or a business

Can we define this a little more? I am a video editor. I have never had ECC RAM. It's never seemed to be a problem? And if I encoded a video file and had a memory era, my encodes at most tend to take 60 minutes. So I'd lose... an hour of a render?

By "critical for work or a business" isn't that again contigent specifically on what the work/business is, IE a render farm at news station with unmissable tight turn deadlines where a corrupt render would be catastrophic VS a professional wedding photographer? (Since we are only talking about live data errors).

 

[Markfw]

Can we define this a little more? I am a video editor. I have never had ECC RAM. It's never seemed to be a problem? And if I encoded a video file and had a memory era, my encodes at most tend to take 60 minutes. So I'd lose... an hour of a render?

By "critical for work or a business" isn't that again contigent specifically on what the work/business is, IE a render farm at news station with unmissable tight turn deadlines where a corrupt render would be catastrophic VS a professional wedding photographer? (Since we are only talking about live data errors).

I will let somebody else answer here. I am not sure what happens when a video gets corrupted. A little bad pixel ? I have no idea.

But here are some other scenarios I do know about. Banking, any error can cause huge problems. Medical, corrupted medical records ? I could go on. Any server that serves a business probably has one or more critical systems like these. A workstation for an accountant ? Same situation.

 

[LikeLinus]

It could be anything from dropped frames, failed renders, and unplayable (corrupt) video file. We use ECC memory though, so I don't typically run into these issues. After Effects will occasionally fail due to a plug-in that has an issue, but otherwise they run rock solid.

Engineering (whatever software they use), 3D, AutoCAD and such would be workstation based usage as well.

 

[scannall]

Can we define this a little more? I am a video editor. I have never had ECC RAM. It's never seemed to be a problem? And if I encoded a video file and had a memory era, my encodes at most tend to take 60 minutes. So I'd lose... an hour of a render?

By "critical for work or a business" isn't that again contigent specifically on what the work/business is, IE a render farm at news station with unmissable tight turn deadlines where a corrupt render would be catastrophic VS a professional wedding photographer? (Since we are only talking about live data errors).

If you were unaware of memory errors for a while, even in video it would be bad. Delivering a client a video that doesn't play correctly or on their hardware, drops frames or artifacts all over for instance would be bad for business.

 

[dsc106]

If you were unaware of memory errors for a while, even in video it would be bad. Delivering a client a video that doesn't play correctly or on their hardware, drops frames or artifacts all over for instance would be bad for business.

I agree, but I've never had this on non ECC memory in 20 years. It's hard to think I've just gotten lucky?

Perhaps there's been a time slipping my mind when something rendered weird and I redid it and it auto-magically fixed itself, and that might be attributed to a non ECC memory error?

Engineering (whatever software they use), 3D, AutoCAD and such would be workstation based usage as well.

I promise I am not trying to be difficult, I am just trying to understand and qualify this. The banking records and database issues that someone else mentioned make sense when running a business background, but here we have content creator tasks like 3D and AutoCad which if I am not mistaken offer perform better on many HEDT processors. Why would this be "workstation use"? (keeping in mind that workstation is not equivalent to ECC memory as we have already established). What does that even mean? What specific features make a workstation better, especially when they are getting slower performance?

 

[IntelUser2000]

And, even for enterprise, how much merit there *actually* is to them over HEDT processors?

I get your point for "prosumers" but for enterprise that's kinda crazy.

- As multicore has expanded on HEDT, benefits of workstation decline (when most software can't even scale to as many threads as a single chip can do, multi-CPU support is extremely niche need)

For most content creation that might be true. For enterprise there are many that require lot of cores and threads.

- ECC memory reliability is such a minor improvement it only matters in server scenarios or really complex calculations

Again, true to some extent for client-focused content creators. For enterprise they opt for multiple 9's uptime(that is, 99.9999....%). They not only have ECC for system memory, but things such as:
-SDCC, which allows recovery in case of a single chip(like in a memory stick) failure.
-Memory Mirroring: If you have two sticks per channel, one is a mirror
-Memory rank sparing: Reserves some capacity for possible corruption
-The CPU itself often has extensive RAS features like ECC in the memory controllers, the L3 cache, and the Home Agent(does some routing of traffic on the CPU.

- More lanes and I/O hardly matters anymore with so much I/O available on Threadripper PCI-E 4 systems, and by virtue of how flexible and how much data even a single I/O port carries today

Never heard of storage or network servers? They have lot of I/O for storage. And what about multiple graphics systems? FPGAs? Deep Learning accelerators?

- More RAM support is becoming increasingly niche as we reach 128gb-256gb of RAM on desktop

This is also false. They are useful for doing large renders on the fly. Transactional processing has insatiable memory requirements as well. New persistent memory technologies like Optane PMMs allow previously not doable things possible so you'll see more usage scenarios such as in-memory databases.

Specially for memory, if some magical memory comes out tomorrow that allows 20x the capacity, immediately it'll be used. What happens is code that used to be written in a way so its mostly in storage would be changed so its mostly in RAM, since if you can do so, RAM is inconceivably faster.

- More memory channels (6 vs 4 on HEDT or 2 on Desktop) - but in this that big of a benefit or are the benefits miniscule?

True for some client systems. But there's definitely a big part of servers where it can never have enough bandwidth. Even HBM2 limits its capabilities.

Does workstation CPU have a place outside server, enterprise, and long-form scientific calculations like quantum models and that sort of thing? And in large organizations/enterprise, are benefits of the platform becoming increasingly dubious and relying more on managers not wanting to stick there necks out to break status quo under the false impression that a workstation cpu will be "more reliable"?

It is more reliable, if you are hosting a large website with millions of visitors a day. And since the hardware and memory gets hammered more with traffic, things like ECC becomes even more important. Even a day offline may result in profits swinging massively. Is it "crazy" to spend $1 million more on hardware if the business is in the order of billions?

 

[LikeLinus]

"if I am not mistaken offer perform better on many HEDT processors."

"What specific features make a workstation better, especially when they are getting slower performance?

Do you have any data to show a single HEDT is faster that comparable dual xeons?

You can get a Xeon that is 28 core (56 threads) that can boost to 4.4ghz. I'm not sure where you are getting this "slower performance" from. You need to prove that before asking us why that is.

Workstation has a meaning the same that HEDT has a meaning. Use Google as instead of having us explain everything to you. There is a wealth of information available. I'm sure Intel has Xeon product pages that explain how they are used and why to use them.